A potential method for sex estimation of human skeletons using deep learning and three-dimensional surface scanning.

Deep learning based on radiological methods has attracted considerable attention in forensic anthropology because of its superior classification capacities over human experts. However, radiological instruments are limited in their nature of high cost and immobility. Here, we integrated a deep learning algorithm and three-dimensional (3D) surface scanning technique into a portable system for pelvic sex estimation. Briefly, the images of the ventral pubis (VP), dorsal pubis (DP), and greater sciatic notch (GSN) were cropped from virtual pelvic samples reconstructed from CT scans of 1000 individuals; 80% of them were used to train and internally evaluate convolutional neural networks (CNNs) that were then evaluated externally with the remaining samples. An additional 105 real pelvises were documented virtually with a handheld 3D surface scanner, and the corresponding snapshots of the VP, DP, and GSN were predicted by the trained CNN models. The CNN models achieved excellent performance in the external testing using CT-based images, with accuracies of 98.0%, 98.5%, and 94.0% for VP, DP, and GSN, respectively. When the CT-based models were applied to 3D scanning images, they obtained satisfactory accuracies above 95% on the VP and DP images compared to the GSN with 73.3%. In a single-blind trial, a multiple design that combined the three CNN models yielded a superior accuracy of 97.1% with 3D surface scanning images over two anthropologists. Our study demonstrates the great potential of deep learning and 3D surface scanning for rapid and accurate sex estimation of skeletal remains.